Presenting Large Language Models as Companions Affects What Mental Capacities People Attribute to Them

TL;DR

The paper investigates how public messaging about large language models (LLMs) influences beliefs about their mental capacities and subsequent reliance on their outputs. Across two preregistered experiments (N=470 and N=604, nine months apart), framing LLMs as companions boosts beliefs that LLMs possess cognitive and emotional capacities, with more modest effects for machines and tools on beliefs and attitudes. A follow-up QA task with a fictitious LLM (Theta) shows that companion framing’s effect on beliefs generalizes to new samples but can be moderated when users engage with LLM outputs; machine framing reduces reliance when responses are inconsistent. The findings highlight that public discourse about AI affects not only anthropomorphism but also how people interact with AI in information-seeking tasks, underscoring ethical considerations for AI communication and literacy efforts.

Abstract

How does messaging about about large language models (LLMs) in public discourse influence the way people think about and interact with these models? To answer this question, we randomly assigned participants (N = 470) to watch a short informational video presenting LLMs as either machines, tools, or companions -- or to watch no video. We then assessed how strongly they believed LLMs to possess various mental capacities, such as the ability have intentions or remember things. We found that participants who watched the companion video reported believing that LLMs more fully possessed these capacities than did participants in other groups. In a follow-up study (N = 604), we replicated these findings and found nuanced effects on how these videos impact people's reliance on LLM-generated responses when seeking out factual information. Together, these studies highlight the impact of messaging about AI -- beyond technical advances in AI -- to generate broad societal impact.

Figures (5)

• Figure 1: Overview of Study 1. Top: Participants were randomly assigned to watch one of three short informational videos or to watch no video. Bottom: The survey recorded participants' attributions of 40 mental capacities to LLMs and five of their attitudes towards LLMs: human-likeness, trust in the outputs of, confidence in using, confidence in programming, and overall feelings towards LLMs.
• Figure 2: Study 1 results showing estimated marginal means and 95% confidence intervals from statistical analyses. All items were measured using 7-point Likert scales. A: Participants' attributions of mental capacities to LLMs. Leftmost facet shows attributions averaged across all 40 capacities. Right three facets show attributions partitioned by the three exploratory factor analysis categories, which we called emotional, cognitive, and physiological. Open circles indicate capacities that were referenced in the companion video. B: Participants' responses of human-likeness of, trust in, confidence in using, confidence in programming, and feelings towards LLMs.
• Figure 3: Overview of Study 2. Top: Simplified examples of Theta's 4 response types: correct-/consistent, correct-/inconsistent, incorrect-/consistent, and incorrect-/inconsistent. Inconsistencies are underlined. Bottom: Screenshots of the factual question-answering task adopted from kim2025fostering. Participants completed eight of these tasks in addition to completing a survey reporting attributions of 10 mental capacities to LLMs.
• Figure 4: Study 2 results showing estimated marginal means and 95% confidence intervals from statistical analyses. A: Study 2 participants in August 2025 (present) attributed less cognitive capacities to LLMs than Study 1 participants in November 2024 (past), but the increased mean attribution rating when presenting LLMs as companions persisted across these two timepoints. B: Using responses from an LLM to complete a task could reduce participants' attributions of cognitive mental capacities to LLMs for those who watched videos presenting LLMs as tools or companions. C: Proportion of participants' submitted answers that agreed with Theta's answer for each video condition and when Theta's response was consistent or inconsistent. Participants who watched LLMs presented as machines (dark blue) agreed with Theta's inconsistent responses even less than other participants.
• Figure 5: Screenshot of mental capacity attribution survey. The item is in bold and participants select a box from 1-7. Participants can see their progress and must answer every item.